Deep Learning-Derived High-Level Neuroimaging Features Predict Clinical Outcomes for Large Vessel Occlusion

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate neuroimaging biomarkers are crucial for guiding mechanical thrombectomy in large vessel occlusion (LVO) stroke.
• Assessing brain tissue changes pre-treatment aids in determining the best therapeutic strategies.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting clinical outcomes in LVO patients undergoing mechanical thrombectomy.
• To compare the predictive performance of the deep learning model against standard neuroimaging biomarkers.

Main Methods:

• A convolutional neural network (CNN) model with an encoder-decoder architecture was designed for ischemic lesion segmentation and feature extraction from diffusion-weighted images.
• The model was trained and validated on a multicenter retrospective cohort of 250 patients and externally validated on 74 patients.
• Performance was assessed by comparing the area under the receiver operating characteristic curve (AUC) with the Alberta Stroke Program Early CT Score (ASPECTS) and ischemic core volume.

Main Results:

• The CNN model achieved a superior AUC of 0.81±0.06, outperforming ASPECTS (0.63±0.05) and ischemic core volume (0.64±0.05) in predicting good clinical outcomes (modified Rankin Scale 0-2 at 90 days).
• External validation confirmed the significantly superior performance of the deep learning model.
• The model extracted high-level imaging features that provided greater prognostic information than traditional biomarkers.

Conclusions:

• Deep learning-derived imaging features from pretreatment diffusion-weighted images offer a more effective prognostic biomarker for LVO patients undergoing mechanical thrombectomy.
• This AI-driven approach demonstrates potential to enhance clinical decision-making beyond current standard neuroimaging biomarkers.
• Further prospective studies are warranted to confirm the clinical utility of this deep learning model.